To eat, or not to eat

Calorie Counter is an application for analyzing food behavior & calories which endgoal is to bringing more consciousness into food habits. Github page

Role: Owner & Lead developer

Show demo

Features

Developed Computer Vision models & applied them efficiently

Built with FastAPI, deployed with Docker and GCP

Technologies backed up by Scientific Paper

Food Recognition Problem

Despite the presence of dozens of applications, algorithms and systems, the problem of recognizing dishes / food products and their ingredients has not been completely solved by the expert communities of machine learning and computer vision. The main current limitation of health and nutrition tracking apps is the need to manually enter each meal.

Dishes similarity

Neural network approaches from researchers and leaders in the calorie counting segment make it easier to track food intake based on image classification. The systems recognize the class of the object, and as a result, they give the user its caloric content and characteristics. The difficulty in solving this problem is the number of classes of dishes and their similarity in some cases

How many calories are on my plate?

However, this recognition does not answer the main question: “How many calories are on my plate?”. This work is devoted to its deep study and solution. Speaking of deep learning for solving classification and segmentation problems, training the latter requires an input mask of an object with the corresponding class, when the classifier is only a class. By using the Image classification + Salient Object Detection models, comparable accuracy in object mask detection and class recognition is demonstrated. This approach is also applicable for training the segmentation model on a large cluster of noticeable objects and performs similar functions.